• Aucun résultat trouvé

DP: Dynamic Prepage in Postcopy Migration for Fixed-Size Data Load

N/A
N/A
Protected

Academic year: 2021

Partager "DP: Dynamic Prepage in Postcopy Migration for Fixed-Size Data Load"

Copied!
5
0
0

Texte intégral

(1)

HAL Id: hal-01403132

https://hal.inria.fr/hal-01403132

Submitted on 25 Nov 2016

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Distributed under a Creative Commons Attribution| 4.0 International License

DP: Dynamic Prepage in Postcopy Migration for Fixed-Size Data Load

Shuang Wu, Ce Yang, Jianhai Chen, Qinming He, Bei Wang

To cite this version:

Shuang Wu, Ce Yang, Jianhai Chen, Qinming He, Bei Wang. DP: Dynamic Prepage in Postcopy Mi- gration for Fixed-Size Data Load. 11th IFIP International Conference on Network and Parallel Com- puting (NPC), Sep 2014, Ilan, Taiwan. pp.530-533, �10.1007/978-3-662-44917-2_45�. �hal-01403132�

(2)

adfa, p. 1, 2011.

© Springer-Verlag Berlin Heidelberg 2011

DP: Dynamic Prepage in Postcopy Migration for Fixed-Size Data Load

Shuang Wu, Ce Yang, Jianhai Chen, Qinming He, Bei Wang

College of Computer Science, Zhejiang University, Zheda Rd. 38, Hangzhou 310027, China {catting, yvxiang, chenjh919, hqm, wangbei}@zju.edu.cn

Abstract. Postcopy migration is a mature technology in virtualization. However the performance of postcopy is not stable. We find many memory intensive loads having a high proportion of independent fixed-size data (FSD) cases. To improve migration performance, we present DP: an algorithm which applies to intelligently tackle FSD load during postcopy migration. We implement DP as an online algorithm triggered by remote paging, and adjusting to prepage the most appropriate amount of pages related to recent page fault records. DP also has a threshold processing mechanism to prevent from noise which is derived from load size fluctuation. The experimental results show that DP algorithm can significantly reduce response time and implicit downtime in postcopy migration, with an high improvement on QoS.

1 Postcopy Migration

Postcopy migration[1,2] is one of the common methods in VM live migration. It forces to move the running VM to target side, then fetches pages from source side when page fault occurs, and pushes the remaining pages to target side at the same time.

In this paper we target to improve migration performance by reducing the response time for each page fault in postcopy migration, using a prepage method. Response time is bottleneck for migration. It is the time waiting for fetching pages from source side when page fault occurs. A proper prepage method can fetch a proper amount of pages in advance, and reduce the response time for each case.

2 Fixed-Size Data

We build up the model for FSD load. It has many independent cases and every case needs a certain amount of memory pages N to complete its work. If we fetch enough pages (more than N), page fault will not occur during the current case.

In practical situations, some cases in FSD-like load need a different amount of pages, which we called noises. Generalized FSD load should have a low NoiseRate.

There are many domains in real benchmark which can be magnified as fig.1. below.

These domains show the characteristic of FSD with some noises.

(3)

Fig. 1. The FSD-like domains. Page fault addresses are placed in a narrow range.

3 Dynamic Prepage

The purpose of DP is to fetch the most suitable amount of pages before next page fault in a noisy environment. We set a range for guessing from NMin to NMax, try to include the best amount N. Every time we try to fetch an estimated value which called NTest, we should decide which range it belongs to, and modify the limit value NMin and NMax cautiously.

Step 1: We divide the guess result into two situations. If the current page fault address is continuous with the pages we fetched the last time, then they belong to the same case, expressed as (1) below. And the opposite situation expressed as (2).

Addrs AddrsLast NLast

Addrs



 

(1) (2) A guessing result NTest is smaller than N or larger than N. The smaller one goes into step 2, and the larger one jump into step 3.

Step 2: From step 1 we get the conclusion NTest < N. But now we cannot trust this conclusion immediately. The modify of NMin should be very cautious, because we need a reliable value to make sure the subsequent guess is not out of range.

We introduce a new array MinRecord[5] to save the recent five NTest, and MinHit for the continuous times of going into step 2.

When MinHit reaches five, we choose the smallest value in MinRecord[5] as NMin.

If noises in record is smaller than N, they will not affect the correct conclusion.

Otherwise they must be larger than N. We avoid these noises by choosing the smallest one in record. Only noises are continuous for five times can lead us to wrong conclusion. Actually the rate of misguidance is related to NoiseRate and continuous times we choose to refresh NMin as expression 3.

Ratemisguidance = NoiseRatecontinuous times. (3)

When continuous times is 5, we can get a low rate of misguidance about 3.2%, even though NoiseRate is up to 50%. This makes the modify of NMin credible.

Then we jump into step 4.

(4)

Step 3: Similar to step 2, we modify NMax cautiously. An array MaxRecord[5] to keep records of the recent five NTest when it placed in the range between N and NMax. MaxHit saves the continuous times of going into step 3. When MaxHit reaches five we choose the largest one in MaxRecord[5]to be NMax.

Every time executing step 3, we should reset MinHit to zero. Similarly when starting from step 2 we should reset MaxHit to zero.

Then the program jump into step 5.

Step 4: Our purpose is to rapidly narrow the gap between NMin and NMax, and also correctly. So we adjust our guessing value NTest according to the current MinHit.

We fetch a smaller amount of pages when MinHit is close to the predetermined threshold. Otherwise a large adjustment as equation 4.

2

NMax NTestlast

NLast

MinHit

 

 . (4)

NLast is the amount we fetch this time and as a parameter transferred to the next round. NTestlast is the value last round we guessed. So that NTest can be calculated as equation 5.

NTestNTestlastNLast. (5) Then we get the amount of pages and return to Step 1.

Step 5: This time we try to rapidly narrow the gap from the NMax side.

We fetch the amount of pages closer to NMax when MaxHit is close to the predetermined threshold. Otherwise a further address from NMax as equation 6.

2

last last

NTest NMin NTest NTest

MaxHit

  

 . (6)

Then set the NLast equal to NTest and return to Step 1.

Strategy of guessing process is as fig. 2. below. MinHit and MaxHit decide the next position of our guess. The purpose of recording hit times is to rapidly and reliably modify the limit both minimum side and maximum side. Each time going into Step 4 or Step 5, the opposite hit times will reset to zero.

Last Page Fault

NMin N NMax

MinHit=5 MinHit=1

MaxHit=1 MaxHit=5

Address

Fig. 2. The way adjusting NTest in both smaller side and larger side. Also the amount that we fetch is relating to MinHit or MaxHit. In this figure we see the different position of next

guessing when MinHit = 1 and MinHit = 5. The same way that MaxHit works.

(5)

The complexity of DP algorithm is O(N), N is for the rounds of page fault in the whole migration.

4 Experimental Evaluation

We use simulation and benchmark ways to evaluate DP algorithm. In simulation experiment DP algorithm works well with the deviation below 0.05 when noisy rate is lower than 20 as fig.3. below.

Fig. 3. Noisy rate and the result of NMin/N and NMax/N after DP algrithm in simulation experiment. Vertical axis shows the result of NMin and NMax, N is always set to be 1.

Horizontal axis means the noisy rate for one case.

We add DP algorithm into postcopy migration and test the new migration on QEMU1.4 [3], Xeon CPU 2.13GHz, 16GB memory and gigabit network. In benchmark way we use STREAM [4] which is a FSD-like benchmark. And DP algorithm can reduce 33% response time in benchmark test.

Acknowledgments. This research is partly supported by the key Science and Technology Innovation Team Fund of Zhejiang under Grant~(No.\ 2010R50041).

5 References

1. Hines, M.R., Deshpande, U., Gopalan, K.: Post-copy live migration of virtual machines.

ACM SIGOPS Operating Systems Review 43(3), 14–26 (2009)

2. Hirofuchi T, Nakada H, Itoh S, et al.: Reactive consolidation of virtual machines enabled by postcopy live migration. In: Proceedings of the 5th international workshop on Virtualization technologies in distributed computing. ACM, 11-18 (2011).

3. McCalpin, John D.: STREAM: Sustainable memory bandwidth in high performance computers. Silicon Graphics Inc (1995).

4. QEMU, http://www.qemu.org/

Références

Documents relatifs

Cognitive scientists shed new light on core concepts of the social sciences: rationality, expectation, identity, social status, intergroup communication, and time perception

The good agreement between our model and observations suggests that motile phytoplankton use mechanical stability to sense the intensity of turbulent cues (Fig. 5A) and integrate

7 grown under the same conditions at a fixed period of the day to avoid diurnal fluctuations in population physiology (S12), and we do not expect the antioxidant capacity to vary

experimentally tested (1 min rolling, which corresponds to 10 full revolutions at Ω = 1 rad s -1 ), by taking the ratio of the stress s for the subpopulation at the bottom (data

The computational domain is limited by an artificial boundary and the numerical waves are then obtained by solving harmonic wave equations coupled with specific boundary con- ditions

For a more advanced analysis, it goes without saying that the general political framework regarding irregular immigration must be put into relation to the legislative framework.

The communication matrix shows that the coordinator node (s5), the redundant task instances (s6,..., s9), and the backup node (s10) communicate time- triggered over reserved

L’épidémiologie des méningites bactériennes aiguës communautaires de l’enfant dans les pays industrialisés a été modifiée en grande partie en raison de la vaccination